morpheus.stages.inference.triton_inference_stage.TritonInferenceStage

Bases: morpheus.stages.inference.inference_stage.InferenceStage

Perform inference with Triton Inference Server.

This class specifies which inference implementation category (Ex: NLP/FIL) is needed for inferencing.

Parameters

c : morpheus.config.Config

Pipeline configuration instance.

model_name

Name of the model specifies which model can handle the inference requests that are sent to Triton inference server.

server_url

Triton server URL.

force_convert_inputs

Instructs the stage to convert the incoming data to the same format that Triton is expecting. If set to False, data will only be converted if it would not result in the loss of data.

use_shared_memory

Whether or not to use CUDA Shared IPC Memory for transferring data to Triton. Using CUDA IPC reduces network transfer time but requires that Morpheus and Triton are located on the same machine.

needs_logits

Determines whether a logits calculation is needed for the value returned by the Triton inference response. If undefined, the value will be inferred based on the pipeline mode, defaulting to True for NLP and False for other modes.

inout_mapping

Dictionary used to map pipeline input/output names to Triton input/output names. Use this if the Morpheus names do not match the model. If undefined, a default mapping will be used based on the pipeline mode as follows:

• FIL: {"output__0": "probs"}

• NLP: {"attention_mask": "input_mask", "output": "probs"}

• All other modes: {}

From the command line this can be specified multiple times for each key/value pair, for example:

–inout-mapping mask input_mask –inout-mapping output probs

which will be inroduced as:

inout_mapping={“mask”: “input_mask”, “output”: “probs”}

Attributes

has_multi_input_ports

Indicates if this stage has multiple input ports.

has_multi_output_ports

Indicates if this stage has multiple output ports.

input_ports

Input ports to this stage.

is_built

Indicates if this stage has been built.

is_pre_built

Indicates if this stage has been built.

name

The name of the stage.

output_ports

Output ports from this stage.

unique_name

Unique name of stage.

Methods

_build(builder, input_nodes)[source]

This function is responsible for constructing this stage’s internal mrc.SegmentObject object. The input of this function contains the returned value from the upstream stage.

The input values are the mrc.Builder for this stage and a list of parent nodes.

Parameters

builder : mrc.Builder

mrc.Builder object for the pipeline. This should be used to construct/attach the internal mrc.SegmentObject.

input_nodes : list[mrc.SegmentObject]

List containing the input mrc.SegmentObject objects.

Returns

list[mrc.SegmentObject]

List of tuples containing the output mrc.SegmentObject object from this stage.

accepted_types()[source]

Accepted input types to this stage.

Returns

typing.Tuple

Tuple of input types.

build(builder, do_propagate=True)[source]

Build this stage.

Parameters

builder : mrc.Builder

MRC segment for this stage.

do_propagate

Whether to propagate to build output stages, by default True.

can_build(check_ports=False)[source]

Determines if all inputs have been built allowing this node to be built.

Parameters

check_ports

Check if we can build based on the input ports, by default False.

Returns

bool

True if we can build, False otherwise.

can_pre_build(check_ports=False)[source]

Determines if all inputs have been built allowing this node to be built.

Parameters

check_ports

Check if we can build based on the input ports, by default False.

Returns

bool

True if we can build, False otherwise.

compute_schema(schema)[source]

Compute the schema for this stage based on the incoming schema from upstream stages.

Incoming schema and type information from upstream stages is available via the schema.input_schemas and schema.input_types properties.

Derived classes need to override this method, can set the output type(s) on schema by calling set_type for all output ports. For example a simple pass-thru stage might perform the following:

If the port types in upstream_schema are incompatible the stage should raise a RuntimeError.

get_all_input_stages()[source]

Get all input stages to this stage.

Returns

list[morpheus.pipeline.pipeline.StageBase]

All input stages.

get_all_inputs()[source]

Get all input senders to this stage.

Returns

list[morpheus.pipeline.pipeline.Sender]

All input senders.

get_all_output_stages()[source]

Get all output stages from this stage.

Returns

list[morpheus.pipeline.pipeline.StageBase]

All output stages.

get_all_outputs()[source]

Get all output receivers from this stage.

Returns

list[morpheus.pipeline.pipeline.Receiver]

All output receivers.

get_needed_columns()[source]

Stages which need to have columns inserted into the dataframe, should populate the self._needed_columns dictionary with mapping of column names to morpheus.common.TypeId. This will ensure that the columns are allocated and populated with null values.

property has_multi_input_ports: bool

Indicates if this stage has multiple input ports.

Returns

bool

True if stage has multiple input ports, False otherwise.

property has_multi_output_ports: bool

Indicates if this stage has multiple output ports.

Returns

bool

True if stage has multiple output ports, False otherwise.

property input_ports: list[morpheus.pipeline.receiver.Receiver]

Input ports to this stage.

Returns

list[morpheus.pipeline.pipeline.Receiver]

Input ports to this stage.

property is_built: bool

Indicates if this stage has been built.

Returns

bool

True if stage is built, False otherwise.

property is_pre_built: bool

Indicates if this stage has been built.

Returns

bool

True if stage is built, False otherwise.

async join()[source]

On all inference worker threads, this function applies join.

property name: str

The name of the stage. Used in logging. Each derived class should override this property with a unique name.

Returns

str

Name of a stage.

property output_ports: list[morpheus.pipeline.sender.Sender]

Output ports from this stage.

Returns

list[morpheus.pipeline.pipeline.Sender]

Output ports from this stage.

async start_async()[source]

This function is called along with on_start during stage initialization. Allows stages to utilize the asyncio loop if needed.

stop()[source]

Stops the inference workers and closes the inference queue.

supports_cpp_node()[source]

Specifies whether this Stage is capable of creating C++ nodes. During the build phase, this value will be combined with CppConfig.get_should_use_cpp() to determine whether or not a C++ node is created. This is an instance method to allow runtime decisions and derived classes to override base implementations.

property unique_name: str

Unique name of stage. Generated by appending stage id to stage name.

Returns

str

Unique name of stage.